复杂背景图像中人体轮廓的自动提取  

Automatic Extraction of Human Body Contours in Complex Background Images

作　　者：林瑞冰 罗芊芊 葛苏敏 吴卓俊 徐平华 LIN Ruibing;LUO Qianqian;GE Sumin;WU Zhuojun;XU Pinghua(School of Fashion Design&Engineering,Zhejiang Sci-Tech University,Hangzhou,Zhejiang 310018,China;Digital Intelligence Style and Creative Design Research Center,Key Research Center of Philosophy and Social Sciences,Zhejiang Province,Zhejiang Sci-Tech University,Hangzhou,Zhejiang 310018,China;Key Laboratory of Silk Culture Heritage and Products Design Digital Technology,Ministry of Culture and Tourism,P.R.China,Hangzhou,Zhejiang 310018,China)

机构地区：[1]浙江理工大学服装学院,浙江杭州310018 [2]浙江省哲学社会科学重点培育研究基地浙江理工大学数智风格与创意设计研究中心,浙江杭州310018 [3]丝绸文化传承与产品设计数字化技术文化和旅游部重点实验室,浙江杭州310018

出　　处：《北京服装学院学报（自然科学版）》2024年第2期104-110,共7页Journal of Beijing Institute of Fashion Technology：Natural Science Edition

基　　金：浙江省哲学社会科学规划交叉学科及冷门“绝学”课题(24LMJX09YB);浙江省重点研发计划项目(2024C01210);浙江省研究生教育学会科研项目(2023-012);浙江理工大学基本科研业务费“科研发展专项”(24076109-Y);浙江省大学生科技创新活动计划暨新苗人才计划(2023R406072);国家级大学生创新创业训练计划项目(202310338047,202210338019)。

摘　　要：为准确提取人体轮廓,提出基于MINet模型的复杂背景人体轮廓提取方法。拍摄正、侧面人体全身照并标注掩膜图,匹配不同的复杂背景,融合形成2860张多场景下的人像数据集。利用迁移学习机制,优化MINet显著目标检测模型,提取人体轮廓。分别对比了基于迁移学习的策略与原模型、U2Net显著目标检测模型、Mediapipe人体轮廓提取算法和传统阈值分割算法的人体轮廓提取效果。结果表明:基于迁移学习的MINet模型具有最优的人体轮廓提取性能,其准确率、精度、召回率和综合指标分别达到了0.998、0.987、0.992和0.990,人体轮廓提取效果与标注的掩模图最为近似。该方法能低成本、批量化、快速提取图像中的人体轮廓,为远程服装定制中的照片测量提供有效的技术方法。To accurately extract human body contours,a method for complex background human body contour extraction based on the MINet model was proposed.Front-view and side-view human images were captured,and corresponding mask images were annotated.Diverse complex backgrounds were matched with portraits to create a dataset of 2860 portraits across various scenes.Using a transfer learning mechanism,the MINet salient object detection model was optimized to extract human body contours.The human contour extraction effects of the transfer learn-ing-based MINet model were compared with the original model,the U2Net salient object detection model,the Mediapipe human contour extraction algorithm,and a traditional threshold-based segmentation algorithm.The results show that the transfer learning-based MINet model demonstrates optimal performance in human body contour extraction,with precision,accuracy,recall,and the composite metric F 1 reaching 0.998,0.987,0.992,and 0.990,respectively,closely resembling the annotated mask images.This method offers a cost-effective,scalable,and fast approach to extract human body contours from images,providing an effective technique for photo measurement in remote clothing customization.

关 键 词：人体轮廓提取 MINet模型 人体尺寸测量 远程测体技术 

分 类 号：TS941.17[轻工技术与工程—服装设计与工程]